There is a large disparity between composite images based on image registration. The composite images may have significantly different appearances due to intensity variation of the images used to generate the composite images. For example, satellite and aerial images taken of the same object may have markedly different intensities, at different times of day, different seasons of the year, different image angles, different altitudes, different image captures systems, or the like. In some fields such as medical imaging or biometrics, image alignment is critical for diagnosis and identification. However, images of the same object when examined at a detailed level may not be properly aligned due to intrinsic and extrinsic variation.
Image registration aims to find the geometrical transformation to align two or more images into the same coordinate system. Geometric transformations have typically been estimated as rigid, piecewise-rigid, or non-rigid. Non-rigid is the most complex. Some existing methods of image registration can be classified based on the variables used in non-rigid registration, into feature-based registration and intensity-based registration.
Many non-rigid techniques have been proposed, most of which are based on minimizing an energy function containing a distance, or similarity, measure and a regularization term. The regularization may encourage certain types of transformations related to different applications and the minimum distance may correspond to a correct spatial alignment.
One type of distance measure is based on mutual information (MI) of images. However, in many applications, the intensity fields of the images may differ significantly. For example, slow varying intensity bias fields often exist in brain magnetic resonance images; while in temporal registration of retina images, the images may contain severe intensity artifacts. As a result, many of the existing intensity based distance measures are not robust to these intensity distortions. There are also methods for simultaneous registration and intensity correction, but they may involve much high computational complexity and suffer from multiple local minima.
In another image registration method, a similarity measure called residual complexity (RC) has been used, where the discrete cosine transform (DCT) is used to code the residual image. RC may be more accurate and robust than MI, for dealing with intensity distortions requiring several parameters, such as regularization, to effect the computation.